Image-based personalized facial expression recognition system using fuzzy neural networks

Abstract

Facial expression is the most natural and comfortable way for human-human communication. Many researches are interested in this field not only psychological field but also engineering field during last few decades. In psychological field, Dr. Ekman insisted that there are six universal facial expressions nonetheless race, culture, district, sex and so on. Six facial expressions are happy, sad, fear, disgust, surprise and angry. Although the Ekman``s work is very well-known in this field, it is very difficult for normal person (who has not sufficient experience for making such expressive expressions) to make such expressive expressions. Thus, we need some remedies for this kind of problems. In this thesis, we propose a concept `personalized service`` for this purpose. In recent days, the `personalized service`` is mainly focused on many fields. And we believe that this kind of personalization would be very important and urgently necessary in these days with tremendous and fast information flow. Focusing on the relation between the user and the service agent, we propose a general structure for personalized service and a personalized facial expression recognition system as an example. In view of technical point, we construct the system for personalized facial expression recognition with fuzzy neural networks and the facial image processing. Fuzzy neural net-works have adopted many advantages from fuzzy decision making system and neural net-works. Human expert``s knowledge is easily implemented due to the structure of fuzzy decision rule system and learning function is added on the basis of neural networks. In the facial image processing, total procedure is divided into three steps. At first, in the face segmentation, a novel scheme to extract the face location is proposed using color histogram based adaptive threshold and motion information. To locate each facial component in the facial region, T-shape based deformable template matching and grouping by projection ...